What Is Kimi K3? China’s New AI Model Explained

On July 16, 2026, Beijing-based startup Moonshot AI released Kimi K3, a 2.8-trillion-parameter artificial intelligence model that it describes as the largest open-weight model to date. Independent testing firm Artificial Analysis ranked it fourth out of 189 models on its Intelligence Index, putting it within range of top systems from OpenAI and Anthropic. The release has drawn attention not just for its size, but because it arrives open-weight, meaning developers will eventually be able to download and run it themselves, at a moment when China’s AI labs are under close watch. Moonshot’s own marketing describes K3 as frontier-level. Independent benchmark data offers a more mixed and specific picture, and that gap is where this explainer starts.

Key Takeaways

• Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model from Moonshot AI, activating only 16 of 896 experts per token.
• Independent benchmarking firm Artificial Analysis scores it 57.1 on its Intelligence Index, ranking fourth of 189 tested models, behind Claude Fable 5 and GPT-5.6 Sol, ahead of Claude Opus 4.8.
• It is open-weight, not fully open source: Moonshot has committed to publishing downloadable weights by July 27, 2026, under a Modified MIT-style license, but has not released training data or code.
• API pricing is $3 per million input tokens and $15 per million output tokens, roughly in line with Western frontier pricing rather than undercutting it.
• Kimi K3 posted particularly strong results in coding benchmarks, including the top spot on LMArena’s Frontend Code Arena.
• Moonshot AI was valued at roughly $20 billion in a May 2026 funding round backed by Alibaba, Tencent and Meituan.

What is Kimi K3? Chinese startup Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight AI model, on July 16, 2026. The model uses a mixture-of-experts architecture and targets coding, reasoning, and agentic tasks. It also offers a 1-million-token context window and native image understanding. Independent tests place it fourth among all tested models on general intelligence, just behind the leading systems from OpenAI and Anthropic.

What Is Kimi K3?

Kimi K3 is the newest flagship model from Moonshot AI, a Beijing-based AI lab founded in March 2023 by Yang Zhilin, Zhou Xinyu, and Wu Yuxin. Yang, the company’s CEO, previously worked as a researcher at Google Brain and Meta AI and holds a doctorate from Carnegie Mellon University. Moonshot is best known for its Kimi chatbot and the Kimi K2 family of open-weight models, which built a following in 2025 for strong coding performance and long-context handling.

K3 is a single model rather than a loose collection of variants, though it is offered with different reasoning-effort settings through Moonshot’s API. Architecturally, it is a mixture-of-experts (MoE) system: rather than activating its entire parameter set for every request, it routes each token through a small subset of specialized sub-networks, or “experts.” According to Moonshot’s technical documentation, K3 selects 16 of 896 available experts per token, a small fraction of the total model. The company has not disclosed how many parameters are actively used per token, only the 2.8 trillion total.

Moonshot positions K3 for coding, long-horizon agentic workflows, visual document understanding, and general knowledge work. It supports native image input, a 1,048,576-token context window, and what the company calls an always-on “thinking mode,” in which the model reasons through a problem before producing a final answer rather than switching between reasoning and non-reasoning modes.

What Does the 2.8-Trillion-Parameter Figure Mean?

Parameters are the internal numerical values a model adjusts during training to learn patterns in language, code, and other data. A larger parameter count generally gives a model more capacity to represent complex relationships, but it does not directly translate into better real-world performance.

Two distinctions matter here. Total parameters describe the entire model as stored on disk, all 2.8 trillion values in Kimi K3’s case. Active parameters indicate how many values the model uses to process a single input, making them more relevant to compute cost and, arguably, capability. As a mixture-of-experts model, K3 activates only a small portion of its total parameters for each token. Moonshot uses the same basic approach in Kimi K2, which has 1 trillion total parameters and 32 billion active parameters, while competing systems such as DeepSeek’s roughly 1.6-trillion-parameter V4 Pro follow a similar design. Moonshot has not published K3’s active-parameter count, so direct efficiency comparisons with those models remain incomplete.

Model size is also a poor stand-in for intelligence. Closed models such as GPT-5.6 Sol and Claude Fable 5 have never disclosed their parameter counts at all, which means claims that K3 is “bigger” than its US rivals cannot actually be verified, only that it is the largest model whose builder has published the number. Independent benchmark results, discussed below, are a more reliable gauge of what the model can actually do than the parameter count alone.

What Does Open-Weight Mean?

“Open-weight” and “open source” are often used interchangeably in AI coverage, but they describe different things. An open-weight model makes its trained parameters, the finished, downloadable model file, available for anyone to run, modify, or fine-tune, typically under a specific license. Under the stricter definition from the Open Source Initiative, an open-source model must also provide its training data, training code, and complete reproduction pipeline so others can rebuild it from scratch. A closed, proprietary model like GPT-5.6 Sol or Claude Fable 5 discloses none of this; it is accessible only through a paid API or hosted application controlled by its developer.

Moonshot says it will release Kimi K3’s weights under a Modified MIT-style license by July 27, 2026, following the same approach it used for the Kimi K2 family. The license will likely allow commercial use, self-hosting, and fine-tuning. However, earlier Kimi licenses required products with more than 100 million monthly active users or over $20 million in monthly revenue to display prominent attribution. Moonshot has not published training data or training code for any Kimi model, so K3 does not meet the stricter open-source definition even once its weights ship. At the time of writing, Moonshot had not yet made the weights public. Users could access K3 only through Moonshot’s hosted API, its consumer apps, and third-party platforms such as OpenRouter.

How Strong Is Kimi K3?

Artificial Analysis provides the most detailed independent assessment of Kimi K3 so far. On its Intelligence Index v4.1, which combines nine reasoning, coding, science, and long-context evaluations, K3 scored 57.1 and ranked fourth among 189 model configurations in late July 2026. It trailed Claude Fable 5 and GPT-5.6 Sol but narrowly beat Claude Opus 4.8.

K3 stood out most in coding. It ranked first on LMArena’s Frontend Code Arena with a 1,679 Elo rating, jumping from Kimi K2.6’s 18th-place position. Moonshot also reported strong SWE-bench Verified and Terminal-Bench 2.1 scores, although some results remain independently unverified.

For agentic work, K3 scored 1,668 on GDPval-AA v2, outperforming GPT-5.5 and GLM-5.2 but trailing Fable 5. It was also second on Artificial Analysis’s long-horizon knowledge-work test. However, K3 showed slower output speed and unusually high token usage. Results may vary by prompts, settings, test versions, and hosting conditions across different providers and environments.

Kimi K3 vs GPT and Claude

These figures come from a mix of Moonshot’s own published benchmarks and Artificial Analysis’s independent testing, as noted throughout; where a company has not disclosed a figure (such as a competitor’s exact parameter count or full pricing structure), the table reflects that gap rather than an estimate.

Why Is Silicon Valley Paying Attention?

Kimi K3 lands as part of a broader pattern: Chinese labs, including DeepSeek, Alibaba’s Qwen team and Zhipu (maker of the GLM series), have increasingly released open-weight models that trade a small performance gap for a large cost advantage and freedom from vendor lock-in. Moonshot’s own K2 line built a substantial developer following through 2025 on exactly this trade-off, and Artificial Analysis’s cost-per-task data suggests K3 continues it: at roughly $0.94 per weighted Intelligence Index task, it undercuts Claude Opus 4.8’s $1.80 while running close to GPT-5.6 Sol’s $1.04, even as it remains pricier than lighter open-weight rivals like GLM-5.2 and DeepSeek V4 Pro.

Open weights carry strategic weight beyond price. Companies can download and self-host the model within their own infrastructure, fine-tune it on private data, and adapt it for regulated environments where third-party APIs are not suitable. Closed models such as GPT-5.6 Sol and Claude Fable 5 cannot offer these advantages by design. For developers and smaller AI labs, a high-ranking open-weight model also lowers the barrier to building competitive products without training a frontier system from scratch.

The release also plays into an ongoing policy debate over US export controls on advanced chips to China. Analysts differ on how much those controls have actually slowed Chinese frontier-model development, and Moonshot has not disclosed what hardware it used to train K3. What is verifiable is the competitive signal: a Chinese lab has, for the first time, placed an open-weight model fourth on a leading independent intelligence benchmark, ahead of at least one flagship model from each major US lab.

Has China Closed the AI Gap With the United States?

This is genuinely contested, and the evidence supports more than one reading. Some analysts point to Kimi K3’s Intelligence Index ranking, fourth of 189 models, ahead of Claude Opus 4.8, as concrete proof that the gap between Chinese and US frontier labs has narrowed substantially in under two years. Bank of America analysts, cited in coverage of the launch, characterized the model’s coding and agentic results as a genuine competitive threat rather than a marketing exercise.

Others caution that a single benchmark ranking, however credible its source, does not equal full frontier parity. Claude Fable 5 and GPT-5.6 Sol still outscore K3 on the same index, and both companies are expected to release further updates. Independent testers have not yet reproduced several of Moonshot’s own headline figures, such as its Terminal-Bench and DeepSWE numbers. K3’s output speed also lagged the field median in Artificial Analysis’s testing, a practical constraint that benchmark accuracy scores do not capture. And a benchmark score says nothing about deployment scale, safety research investment, enterprise infrastructure or the deep bench of applications already built on established US model APIs, all factors that shape real-world competitiveness beyond a leaderboard position. Whether Kimi K3 represents a durable shift or a single strong release will become clearer only as OpenAI, Anthropic and Moonshot each ship their next models.

What Are Kimi K3’s Limitations?

Several constraints are worth weighing before treating Kimi K3 as a drop-in replacement for an established frontier model. Its measured output speed, about 62 tokens per second in Artificial Analysis’s testing, trails the roughly 72.7-token median for comparably priced models, which can matter for latency-sensitive applications. The model also used unusually high token counts to reach its benchmark scores, roughly double the peer median, which can increase real-world cost even at a competitive per-token price.

At launch, K3’s open weights were not yet available; Moonshot has committed to a July 27, 2026 release date, but until the actual license file and model card are published, terms for commercial self-hosting remain provisional rather than confirmed. Self-hosting a 2.8-trillion-parameter mixture-of-experts model is also a substantial infrastructure undertaking, Moonshot has recommended dozens of accelerator chips for serving the model at scale, which puts local deployment out of reach for most individual developers regardless of licensing terms. As of this writing, independent evaluators had not reproduced several of Moonshot’s benchmark claims, including specific coding and terminal-agent scores. Readers should therefore treat those figures as company-reported rather than independently confirmed.

Is Kimi K3 Available to the Public?

Kimi K3 is available today through Moonshot’s API, which is compatible with the OpenAI SDK, lowering the integration effort for developers already building on GPT models. It is also accessible through Moonshot’s consumer Kimi apps and through third-party aggregators such as OpenRouter. Pricing on the first-party API is $3 per million input tokens and $15 per million output tokens, with a 90% discount on cached input tokens, bringing that rate to $0.30 per million. Moonshot also offers subscription-based access through its Kimi platform for interactive use, separate from metered API billing.

Downloadable model weights, which would allow self-hosting outside Moonshot’s infrastructure, were not available at launch. The company says it will publish the weights on its Hugging Face organization by July 27, 2026, together with a full technical report. Moonshot has not disclosed any regional restrictions on API access. However, enterprise users in regulated industries should confirm data-residency and compliance terms directly with the company before deploying the model in production.

Why Kimi K3 Matters

For developers, Kimi K3 adds a genuine high-end option to the open-weight category, rather than the mid-tier alternative that open models have typically represented. Businesses evaluating AI vendors may also see its pricing as added pressure on the cost structure of proprietary frontier APIs, particularly for coding and agentic workloads, where K3 delivered its strongest results. More broadly, a downloadable and modifiable model reaching the top five of an independent intelligence ranking establishes a new benchmark against which other labs, both Chinese and Western, will likely be measured. And for the wider US-China technology competition, K3 is another data point suggesting that gap-closing in AI is happening unevenly: faster in some domains, such as coding, than in general reasoning or the deep infrastructure that supports frontier deployment at scale.

Frequently Asked Questions

Q1. Who owns Moonshot AI? 

Moonshot AI is a private, Beijing-based company founded in March 2023 by Yang Zhilin, Zhou Xinyu and Wu Yuxin. Its investors include Alibaba, Tencent, Meituan and China Mobile. The company was valued at roughly $20 billion in a funding round in May 2026 and has reportedly explored a Hong Kong listing.

Q2. Is Kimi K3 open source? 

Not by the strict definition. It is open-weight, meaning Moonshot plans to release the downloadable model file under a Modified MIT-style license by July 27, 2026. It is not open source in the fuller sense, because Moonshot has not released the training data or training code needed to reproduce the model from scratch.

Q3. Is Kimi K3 better than GPT?

It depends which GPT model and which task. On Artificial Analysis’s Intelligence Index, K3 trails OpenAI’s GPT-5.6 Sol overall, but it beat GPT-5.5 on several coding benchmarks and topped GPT-5.6 on measures like Program Bench and BrowseComp. Neither model leads across every category.

Q4. Can developers run Kimi K3 locally? 

Not yet. The model’s weights had not been released as of publication; Moonshot has committed to a July 27, 2026 release date. Once available, self-hosting will require substantial multi-accelerator infrastructure given the model’s 2.8-trillion-parameter size, even though it is a sparse mixture-of-experts design.

Q5. How much does Kimi K3 cost? 

Moonshot’s first-party API charges $3 per million input tokens and $15 per million output tokens, with cached input discounted to $0.30 per million. Independent cost analysis from Artificial Analysis put its effective price at roughly $0.94 per weighted benchmark task, cheaper than Claude Opus 4.8 but pricier than lighter open-weight models like DeepSeek V4 Pro.

Q6. Is Kimi K3 available outside China? 

Yes. The API and Kimi apps are accessible internationally, and the model is also listed on third-party platforms such as OpenRouter. Moonshot has not disclosed specific regional access restrictions, but enterprise users should verify data-handling terms for their jurisdiction directly with the company.

Q7. Is Kimi K3 safe to use? 

Moonshot has not published a detailed third-party safety evaluation of K3 comparable to those major US labs release for their frontier models. As with any AI system, users should independently verify outputs for factual accuracy, especially in high-stakes use cases, and review the eventual license and model card once published for any usage restrictions.